Topology-Constrained Learning for Efficient Laparoscopic Liver Landmark Detection

Ruize Cui, Jiaan Zhang, Jialun Pei, Kai Wang, Pheng-Ann Heng, Jing Qin

Abstract

Accurate anatomical landmark detection is critical for Augmented Reality (AR) navigation in laparoscopic liver surgery. However, standard deep learning models often fail to preserve the thin, tubular topology of these landmarks, leading to fragmented predictions and noise from surgical artifacts. We present TopoNet, a novel framework that incorporates topological constraints to ensure structural integrity. By utilizing a Snake-CNN dual-path encoder and specialized topological loss functions, TopoNet achieves state-of-the-art accuracy and a 4x improvement in inference speed over existing methods, making it suitable for real-time clinical applications.

Introduction

In laparoscopic liver surgery, surgeons operate through minimally invasive incisions guided by a 2D camera feed. To enhance safety and precision, Augmented Reality (AR) systems overlay 3D preoperative models onto this live video. The success of this overlay depends heavily on the accurate identification of liver landmarks, such as vessels and surface boundaries.

Existing AI models struggle with this task because liver landmarks are typically thin, winding, and tubular. Standard segmentation approaches treat these structures as independent pixels, often resulting in "broken" lines or false positives triggered by surgical instruments, blood, or glare. When the connectivity (topology) of a landmark is lost, the AR navigation system cannot reliably align the 3D map, potentially compromising surgical outcomes. TopoNet addresses these challenges by shifting the learning focus from local pixel-matching to global structural preservation.

Methodology

TopoNet introduces three primary architectural innovations designed to respect the unique anatomy of liver landmarks:

1. Snake-CNN Dual-Path Encoder

The model processes both standard RGB video frames and estimated depth maps. While the RGB path extracts texture and color information, the depth path utilizes Snake Topology Acquisition (STA) blocks. These blocks employ "Dynamic Snake Convolutions," which adapt their receptive fields to follow the slender, winding paths of tubular structures, rather than relying on fixed square kernels.

2. Boundary-aware Topological Fusion (BTF)

Instead of simple concatenation, the BTF module uses an attention mechanism to fuse RGB and depth features. This specifically highlights anatomical boundaries, helping the model distinguish between the liver surface and distracting foreground elements like surgical tools.

3. Topological Constraint Loss

The core of TopoNet’s structural accuracy lies in its dual-loss strategy:

  • Center-line Constraint (clDice): This ensures the "skeleton" of the predicted landmark matches the ground truth. By supervising the connectivity of the center line, the model prevents the fragmented segments common in standard U-Net architectures.
  • Topological Persistence Loss: Leveraging algebraic topology (persistent homology), this loss function generates "barcodes" representing the number of connected components. It mathematically penalizes the model for creating isolated clusters of pixels (noise) that do not belong to the continuous anatomical structure.

Experimental Results

TopoNet was evaluated on two benchmarks: the L3D dataset (1,152 annotated keyframes) and the P2ILF dataset (183 challenging images). The model was compared against 12 state-of-the-art frameworks, including SAMed and D2GPLand.

Quantitative Performance

On the L3D dataset, TopoNet achieved a 65.19% Dice Score (DSC) and 50.56% Intersection over Union (IoU). The most significant metric, however, was the Average Symmetric Surface Distance (Assd), which measures structural precision. TopoNet reduced the Assd error to 28.07 pixels, a massive improvement over the 59.38 pixels recorded by the previous best model, D2GPLand.

Qualitative Analysis and Efficiency

Visual results demonstrate that TopoNet successfully maintains the connectivity of landmarks even in the presence of occlusions. Unlike previous models that produce "floating" false positives on surgical instruments, TopoNet’s topological persistence loss effectively filters out such noise.

Qualitative comparison of liver landmark detection. TopoNet (d) maintains superior topological consistency compared to D2GPLand (c), which suffers from fragmented predictions and false positives (red circles).

Furthermore, TopoNet is highly optimized for the operating room. It achieves an inference speed of 86.43ms per frame, making it approximately 4x faster than D2GPLand and capable of supporting real-time AR overlays.

Discussion and Conclusion

TopoNet demonstrates that connectivity is a vital feature for medical imaging tasks involving slender structures. By supervising the skeleton and connectivity of landmarks through algebraic topology, the network achieves a level of robustness that pixel-wise losses cannot provide.

While the model currently depends on external depth estimation—which can be affected by extreme surgical smoke or glare—it represents a major step toward reliable AR-guided surgery. Future work will focus on integrating TopoNet into a complete real-time navigation pipeline and addressing extreme occlusion scenarios. In summary, TopoNet provides a high-performance, efficient solution for liver landmark detection that respects the fundamental topology of human anatomy.

Summary updated

BibTeX
@inproceedings{cui2025topology,
  author = {Ruize Cui and Jiaan Zhang and Jialun Pei and Kai Wang and Pheng-Ann Heng and Jing Qin},
  title = {{Topology-Constrained Learning for Efficient Laparoscopic Liver Landmark Detection}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2025},
  doi = {10.1007/978-3-032-05127-1\_56},
}
Figure